In order to eliminate the effect of the misalignment error between the sensor array and the sensor system of the magnetic gradient tensor system on the measurement accuracy, a method of theoretically precise calibration between all magnetic sensors and reference platforms, which involves rotating a circle around an arbitrary axis of the system, was proposed. The linear correction model of the sensor system error was constructed using two nonlinear transformations without any mathematical simplification, and the ideal orthogonal output of the reference platform and each sensor was obtained with only 10 sets of measurement data in the same rotation period. By constructing the rotation matrix of the tri-axis heel, pitch, and azimuth transformations of the magnetic sensor, the misalignment error correction model of the arbitrary spatial orientation of the sensors was obtained, and the rotation angle was estimated by the least-squares method. In addition, only three sets of measurement data in the same rotation period was necessary for the alignment of the tensor system. The simulation and experiment show that the accuracy of the simulation parameters estimation was close to 100% in the ideal condition. After the calibration experiment, the output of the sensor showed a high overlapping and coaxiality performance, and the RMSE (root mean square error) of the tensor components was less than 30 nT/m. It is possible to improve the measurement accuracy of the differential magnetic gradient tensor system efficiently with simpler steps and less sampling data.